DE60031050T2 - Segmentation of the border of the prostate in two- and three-dimensional ultrasound images - Google Patents

Abstract

A fast semi-automatic prostate contouring method and apparatus using model-based initialization and an efficient Discrete Dynamic Contour (DDC) for boundary refinement. The user initiates the process of the preferred embodiment by identifying four (4) points on the prostate boundary, thereby scaling and shaping a prostate model, and then the final prostate contour is refined with a DDC. The method of the present invention has particular application during the pre-implant planning phase of a brachytherapy procedure. However, this method also has uses in any phase of dose planning in the brachytherapy procedure or any other therapy approach. <IMAGE>

Description

These This invention relates generally to medical imaging systems, and more particularly to a method and apparatus for prostate and marginal segmentation in two- and three-dimensional ultrasound images.

Prostate Cancer is the most common diagnosed malignant Tumor in men over the age of 50 years and is autopsied at 30% of men over the age of 50, at 40% at the age of 60, and at nearly 90% at the age of 90 Found years ago. He is the second cause of death worldwide of cancer in men and accounts for between 2.1% and 15.2% of all cancer deaths. each Year in Canada by this disease will be about 20,000 new prostate cancer cases diagnosed and it dies about 4,000 men.

In general, symptoms due to prostate cancer do not appear until there is extensive local growth or until metastases develop, which is explained as the reason that only 65% of patients with locally limited disease are diagnosed. Once the tumor has spread across the prostate, the risk of metastasis increases dramatically. Tumors smaller than 1 to 1.5 cm ³ rarely break the prostatic capsule. If diagnosed at this early stage, the disease is curable, and even at later stages treatment can be effective. However, the treatment options vary depending on the extent of the cancer, and the prognosis worsens when the diagnosis is made at an advanced stage.

The Challenges faced by doctors faced, the patients with possible Treating prostate cancer are: (a) clinically relevant cancers to diagnose at a stage when they are curable, (b) the disease is precise to classify and classify, (c) to apply the appropriate therapy exactly, about the destruction to optimize the cancer cells while the adjacent normal tissues are preserved; (d) the patients to observe for side effects and the effectiveness of the therapy to judge.

The U.S. Patents 5,562,095, 5,454,371 and 5,842,473 focus on the challenges (a) and (b). These patents describe one 3D ultrasound imaging technique for the diagnosis of prostate cancer. Extensions of these concepts for prostate cryosurgery are described in the commonly owned patent application 60 / 321,049, the contents of which are incorporated herein by reference. In another Application is described a 3D ultrasound-guided intra-operative brachytherapy system. An important consideration in stopping appropriate prostate therapy is the exact segmentation (i.e., working out) of the edge of the prostate and other anatomical structures (rectal wall, urethra, bladder). The assignment of the appropriate therapy or dose for the prostate requires that the prostate volume is measured accurately.

at Conditions with a high image contrast (eg in liquid-filled areas in ultrasound images), the segmentation task is relatively easy and it can many approaches are used. Ultrasound pictures of the However, prostate are very difficult to segment because of the contrast is very low and the picture under spots, shadows and others Artifacts suffers. When performing of ultrasound image segmentation are conventional local image processing operators like edge detectors in and of themselves due to spots, Shadow and other image artifacts to find the edge inadequate.

The Variation in measurements of prostate volume, which is a conventional Using 2D technique is high because of the current ultrasonic volume measurement techniques emanating from an idealized elliptical shape and just simple Use width measurements in two views (see Tong S, Cardinal HN, McLoughlin RF, Downey DB, Window A, "Intra- and inter-observer variability and reliability of prostate volume measurement via 2D and 3D ultrasound imaging ", Ultrasound in Med & Biol 1998; 24: 673-681, and Elliot TL, Downey DB, Tong S, Mclean CA, Window A, "Accuracy of prostate volume measurements in vitro using three-dimensional ultrasound ", Acad Radiol 1996; 3: 401-406).

Manually Contour sequential 2D cross-sectional CT or TRUS (transrectal ultrasound) prostate images has reduced this fluctuation, but this approach is Time-consuming and cumbersome, what to do during an intra-operative treatment makes impracticable.

Domain Neural Network applications require extensive learning sets, are slow, and make the addition of operator-specified edge information difficult. Contour-related procedure Such as snakes' implementation of active contours are slow, complex, and sensitive to the initial choice of the contour.

US Patent No. 5,559,901 to Lobreg discloses a method of determination a shape contour, wherein vertices along the contour of the shape with be bridged straight sections, an estimated To provide initial contour. To smooth the contour is an internal force or energy at each of these vertices of the estimated initial contour Defines the force or energy of the angle between depends on the edges along a sequence of adjacent vertices.

According to the present The invention provides a method of object segmentation according to the independent claims posed. Some optional features are set forth in the dependent claims.

According to one embodiment becomes a fast semi-automatic prostate contouring procedure provided a model-based initialization and a effective Discrete Dynamic Contour (DDC) for edge refinement used. The operator initiates the process of the preferred embodiment Identification of four (4) points on the edge of the prostate, which scales and shapes a prostate model, and then becomes the final one Prostate contour refined with a DDC.

The Method of the present invention may be a 2D segmentation technique include or alternatively until the segmentation of objects expand in 3D.

The Method of the present invention has a specific application the pre-implantation planning phase of a brachytherapy application serves. This method is also used in every phase the dosing schedule in the brachytherapy application or any other therapy application.

The The present invention can be practiced in various ways be implemented. As an example, various embodiments of the present invention with reference to the drawings described, wherein:

1 A prior art segmentation technique using DDC in which 1 (a) an initial DDC drawn by the operator shows 1 (b) shows an early stage of deformation, 1 (c) shows a later stage of reshaping, and 1 (d) the final deformed DDC shows.

2 shows an example of the prior art configuration of DDC.

3 Forces as vector sizes for the configuration of 2 shows that have both direction and size.

4 Shine energy field and apparent force field for a vertical edge of an ultrasound image, wherein 4 (a) showing the edge modeled as a step edge (ie, an edge where there is a drastic transition of grayscale), 4 (b) the energy field associated with the edge shows, and 4 (c) the apparent field shows.

5 is an illustration of a curvature vector, c → _i , as a measure of the angle between edges for (a) a large curvature, and (b) a small curvature.

6 Fig. 10 is a flowchart showing implementation of the DDC conversion process as used in the method of the present invention.

7 Fig. 10 is a flowchart for the 2D segmentation algorithm according to the present invention.

8th shows the location of four points selected by the operator designated (1) to (4) on a prostate atlas according to the method of the present invention.

9 A first example of the operation of the prostate segmentation algorithm of the present invention is shown in FIG 9 (a) shows an initial outline with four points selected by the operator, and 9 (b) shows the final segmented contour obtained after forming with DDC.

10 the steps of segmentation and processing in a second example of the operation of the present invention shows in which portions of the DDC are not very well following the prostatic border, where 10 (a) shows an initial outline with four points selected by the operator, 10 (b) showing the DDC after a first transformation, 10 (c) shows three vertices (shown as squares) that are machined and fixed, followed by further reshaping of the DDC, and 10 (d) shows the outline after second transformation.

11 the segmentation of a "difficult" case according to the present invention, wherein 11 (a) shows the initial outline with four points selected by the operator, 11 (b) the DDC after the first transformation shows 11 (c) shows five vertices (shown as squares) that are machined and fixed, followed by further reshaping of the DDC, and 11 (d) shows the outline after second transformation.

12 Figure 12 is a schematic diagram illustrating how the method of selecting sequences of 2D images from the 3D volume is used for the segmentation.

13 shows a sequence of 2D images of the segmented prostate, using the semi-automatic segmentation method extended to 3D. The pictures are parallel to each other, at a distance of one millimeter. Image # 15 was used to initiate the process through the semi-automatic 2D segmentation technique. Following the segmentation of this 2D image, the trailing edge was used to initiate the segmentation of the next image. This process was repeated until the entire prostate was segmented.

As discussed above, DDC is a polyline (ie, a series of points connected by straight sections) that is used to transform a contour to fit features in an image. When the DDC is used to segment an object from an image as originally described by Lobregt and Viergever in "Discrete Dynamic Contour Model", IEEE Trans. Med. Imag. 1995; 14: 12-24, the operator must first draw an approximate outline of the object. An example outline is in 1 (a) shown for a typical 2D US (ultrasound) image of the prostate gland. The DDC is initialized to this outline and automatically reshaped to match the features of interest, in this case consisting of the edges defining the border of the prostate. 1 (b) and 1 (c) show two stages of forming the DDC. It should be noted that dots are automatically added or deleted to the DDC during reshaping to allow better fitting to the edge. 1 (d) shows the final reshaped DDC. In this example, the final shape is good for the prostatic border, except in some areas where artifacts and noise in the image cause parts of the DDC to reform toward and away from the prostatic border.

definiert am Eckpunkt i durch zwei Kantenvektoren, die mit dem Eckpunkt verbunden sind: wobei ||·|| den Betrag eines Vektors anzeigt. Ein lokal radialer Einheitsvektor r ^_i kann aus t ^_i berechnet werden, indem er um B/2 Radianten gedreht wird: r ^_i und t ^_i definieren ein lokales Koordinatensystem, das im Eckpunkt i angeordnet ist.The configuration of the DDC is in 2 shown. As shown, the unit edge vector d ^ _i points from vertex i to vertex i + 1. A locally tangent unit vector, t ^ _i , is

defined at vertex i by two edge vectors connected to vertex: where || · || indicates the amount of a vector. A local radial unit vector r ^ _i can be calculated from t ^ _i by rotating it by B / 2 radians: r ^ _i and t ^ _i define a local coordinate system located at vertex i.

wobei E die "Energie" darstellt, die zu einem Pixel gehört, das die Koordinaten (x^p, y^p) hat, G_Φ ein Gauss'scher Glättungskern ist mit einer charakteristischen Breite von Φ, und I das Bild ist. Der · Operator stellt die Faltung dar, und ∇ ρ ist der Gradient-Operator. Die Energie hat ein lokales Maximum an einer Kante, und die Kräfte, die aus der Energie berechnet werden, dienen dazu, diese Kante zu lokalisieren.The DDC process is based on simple dynamics. A weighted combination of internal ( fρ int i ), ff ρ img i ) and damping ( f ρ d i ) Forces acts on every vertex i of the DDC, resulting in a total force f ρ dead i results: f ρ dead i = w int i f ρ int i + w img i f ρ img i + f ρ d i (3) Where w img i and w int i relative weights for the apparent and internal forces are. It should be noted that these forces are vector quantities that have both magnitude and direction as in 3 shown. For the prostate, the apparent forces cause the vertices to float against the edges. These powers can be for every pixel of the image are defined as: E (x, y) = || ∇ ρ (G Φ · I (x p , y p )) || (4a)

where E represents the "energy" associated with a pixel having the coordinates (x ^p , y ^p ), G _{Φ is} a Gaussian smoothing kernel with a characteristic width of Φ, and I is the image. The operator represents the convolution and ∇ ρ is the gradient operator. The energy has a local maximum at one edge, and the forces that are calculated from the energy serve to locate that edge.

The factor of two in equation (4b) assigns to the magnitude of the apparent force a range of zero to two that is the same as that for the internal force defined below. An example energy field and its associated force field for a simulated image of an edge are in 4 shown. It should be noted that apparent forces have an effect over a limited spatial extent around an edge, such as the concentration of energy around the edge in (4b) shown. The extent of this "detection area" is determined by σ in equation (4a), and becomes larger as σ becomes larger. Although a large detection range is desirable to compensate for operator deficiencies in the initial contour of the DDC (eg, some vertices on the initial DDC may be out of the detection range), a large value of σ may result in poor location of the desired edge , A large σ also offers the advantage of increased noise reduction. According to a successful implementation of the present invention, for prostate images, a value of σ = 5 pixels has been chosen as a compromise.

To evaluate the apparent force acting on the vertex i of the DDC, the force field expressed by Equation (4b) is sampled at the vertex coordinates (x _i , y _i ) using bilinear interpolation. Furthermore, only the radial component of the field acts on the vertex, since the tangential component can potentially cause the vertices to bunch together as the DDC transforms so that the resulting apparent force is at vertex i: ff ρ img i = (f ρ img (x i , + y i ) · R ^ i ) r ^ i (5) where · denotes a scalar product of vectors.

wobei c ρi = d ^i – d ^i-1 (7)der Krümmungsvektor am Eckpunkt i ist. Der Krümmungsvektor ist ein Maß des Winkels zwischen zwei Kanten wie in 5 gezeigt. Obwohl unter der einfachen Annahme, dass f ρinti zu c ρ_i proportional ist, das Ziel der Minimierung der lokalen Krümmung erreicht werden könnte, hat die Definition in Gleichung (6) den Vorteil, zu verhindern, dass die DDC zu einem Punkt bei Abwesenheit von Scheinkräften zusammenfällt.The internal force causes the local curvature at each vertex to be minimized, keeping the DDC smooth in the presence of image noise. It is defined as:

in which c ρ i = d ^ i - d ^ i-1 (7) the curvature vector at vertex i is. The curvature vector is a measure of the angle between two edges as in 5 shown. Although under the simple assumption that f ρ int i to c ρ _{i is} proportional, the goal of minimizing the local curvature could be achieved, the definition in equation (6) has the advantage of preventing the DDC from collapsing to a point in the absence of apparent forces.

The damping force at vertex i is assumed to be proportional to the velocity (v ρ _i ) at this vertex: f ρ d i = w d i v ρ i (8th) in which w d i is a negative weighting factor. The damping force keeps the DDC stable (ie, prevents oscillations) during forming.

To simplify matters, uniform weights have been used for the apparent, internal and damping forces selected for the result that is in 1 is shown, and for all subsequent segmentations. For each vertex, the weights of w img i = 1.0 . w int i = 3.0 and w d i = -0.5 accepted. Although these values were selected experimentally, the greater weighting for the apparent force promotes deformation of the contour to the edges rather than smoothing due to internal forces. If there had been much more noise in the pictures, the internal forces would have been given more weight. The selection of the weighting for the damping force seems to be less critical, and a small negative value has been determined, which provides good stability.

wo p ρ_i die Position des Eckpunktes ist, a ρ_i seine Beschleunigung ist, m_i seine Masse ist und Δt der Zeitschritt für numerische Integration ist. Zur Vereinfachung wird die Masse eines jeden Eckpunkts als Eins angenommen. Die Anfangsposition eines jeden Eckpunkts (d. h. die Position bei t = 0) ist spezifiziert durch die Bedienergezeichnete Kontur, und die Anfangsgeschwindigkeit und Beschleunigung sind auf Null gesetzt. Die Position, Geschwindigkeit und Beschleunigung von jedem Eckpunkt werden dann unter Verwendung der obigen Gleichungen iterativ aktualisiert. Nach jeder Iteration wird die DDC "resampled", um ihr Eckpunkte hinzuzufügen oder aus ihr zu entfernen, wie unten detaillierter beschrieben wird. Die Iterationen werden fortgesetzt, bis alle Eckpunkte zur Ruhe kommen, was geschieht, wenn die Geschwindigkeit und die Beschleunigung von jedem Eckpunkt näherungsweise 0 werden (d. h. wenn ||v ρ_i|| ≤ ε₁ und ||a ρ_i|| ≤ ε₂, wobei ε₁ und ε₂ zwei kleine positive Konstanten nahe Null sind). Ein Zeitschritt von Eins wird für die Integration verwendet.Forming the DDC is controlled by the above forces and is implemented as a simple numerical integration, presented in 6 , The position of the vertex i on the DDC at time t + Δt is calculated from its position at time t using the following equations of the finite difference method: pp ρ i (t + Δt) = p ρ i (t) + vv ρ i (t) Δt (9a) v ρ i (t + Δt) = v ρ i (t) + a ρ i (t) Δt (9b)

where p ρ _{i is} the position of the vertex, a ρ _{i is} its acceleration, m _{i is} its mass, and Δt is the time step for numerical integration. For simplicity, the mass of each vertex is assumed to be one. The initial position of each vertex (ie the position at t = 0) is specified by the operator drawn contour, and the initial velocity and acceleration are set to zero. The position, velocity and acceleration of each vertex are then iteratively updated using the equations above. After each iteration, the DDC is "resampled" to add or remove vertices, as described in more detail below. The iterations continue until all vertices come to rest, which happens when the velocity and acceleration of each vertex becomes approximately 0 (ie, when || v ρ _i || ≤ ε ₁ and || a ρ _i || ≤ ε ₂ , where ε ₁ and ε _{2 are} two small positive constants near zero). A time step of one is used for the integration.

As discussed above, the border of the prostate is at times difficult to define because 2D US images of the prostate suffer from spots, shadows, and refractions. In addition, the contrast between the prostate and the surrounding structures is small and depends on the transducer of the system and the ultrasound frequency. These effects combine and make the selection of the leading edge difficult as they allow the DDC to converge to the correct edge. The vertices on the DDC must be initialized close enough to the desired edge to be drawn to it, rather than to nearby insignificant edges. Exact initialization can require considerable operator effort, as the operator may be required to enter many points, as in 1 (a) shown. To overcome this problem, according to the present invention, information on the prostate-specific form is incorporated in the initialization phase. With this approach, it is possible to get a good estimate of the prostate bone with very little operator input. It is even possible, despite good initialization, for the DDC to converge to strong non-prostate edges that are nearby. The incorporation of gradient direction information into the apparent force field can reduce the attraction of such features.

The algorithm of the present invention consists of three main phases: initialization, transformation and processing. Each process is described in detail below. The interaction of these processes is in 7 shown. After the operator has selected an image and adjusted the window and level, he must initialize the DDC by selecting four points on the edge. The operator can choose to set the four points, that is to make them immobile. The initial DDC is then reformed. If the DDC accurately replicates the prostatic border after it has been reshaped, it can be manipulated by allowing the operator to interactively move and fix vertices. The machined DDC can be further reshaped if necessary.

As mentioned above, the DDC initialization routine of the present invention requires the operator to select only four points named from (1) to (4) in FIG 8th , Points (1) and (3) lie on an approximated axis of symmetry. A local xy coordinate system can be defined by orienting the y-axis from point (1) to point (3). The x-axis is then taken perpendicular to the y-axis and is oriented so that the cross product of the unit vector points out of the screen along the x-axis and the y-axis (ie, x ^ x ŷ). The origin of the coordinate system is centered on the points (1) and (3) set.

wobei s ein Parameter ist, der entlang der Kurve variiert, dann werden die Punkte (2) und (4) auf den "Lappen" der Prostata ausgewählt, wo die lokale Tangente zu der Kurve an diesen Punkten parallel zu der y-Achse ist, d. h. Punkte, mit x'(s) = 0. (11) If the initial form of the DDC is represented in parametric form, such as:

where s is a parameter that varies along the curve, then points (2) and (4) are selected on the "lobe" of the prostate, where the local tangent to the curve at these points is parallel to the y-axis, ie points, with x '(s) = 0. (11)

The 'operator designates Differentiation according to s.

The four points split the prostatic border into four distinct segments: segment 1 to 2, which starts at point 1 and ends at point 2, and segments 2 to 3, 3 to 4, and 4 to 1. The initial shape of the prostate is Hermite's Estimated interpolation of the endpoints to automatically obtain additional points within each segment. It has been discovered that cubic interpolation functions provide a good approximation for a wide range of prostate forms and have the following form: x (s) = a 3 s 3 + a 2 s 2 + a 1 s + a 0 (12a) y (s) = b 3 s 3 + b 2 s 2 + b 1 s + b 0 (12b) where s is a parameter that varies from 0 at the start point of the segment to 1 at the end point and where a _i and b _i (i = 0, 1, 2, 3) are coefficients.

Around additional It's first to interpolate points within a segment necessary, to calculate the coefficients of equation (12). Two data members are required at each endpoint: the value of the variable (either x or y) and the rate of change of these variables with respect to s (x 'or y'). Table 1 summarizes the information the for every segment is required. The table assumes each segment starts at s = 0 and ends at s = 1. Some Points are to be noted.

First, the tangents at points (1) and (3) are approximately parallel to the x-axis, ie at these points we have y '(s) = 0. (13)

Here, the rate of change of x with respect to s is selected from the operator selected points as listed in Table 1, which shows the data required to calculate the coefficients for the interpolation functions for each segment (1-2, 2-3 , 3-4, and 4-1), where (x ₁ , y ₁ ), (x ₂ , y ₂ ), (x ₃ , y ₃ ) and (x ₄ , y ₄ ) refer to the coordinates of the four from Refer to selected points of the operator shown in 3 ,

Second, at points (2) and (4) where equation (11) holds, the rate of change of y with respect to s must also be estimated as listed in the table. With these data, the coefficients can then be calculated using the formulas: a 0 = x (0) a 2 = 3 (x (1) - x (0)) - x '(1) - 2x' (0) a 1 = x '(0) a 3 = 2 (x (0) - x (1)) + x '(0) + x' (1) (14a) and b 0 = y (0) b 2 = 3 (y (1) - y (0)) - y '(1) - 2y' (0) b 1 = y '(0) b 3 = 2 (y (0) -y (1)) + y '(0) + y' (1) (14b)

When the coefficients are then calculated, the points are uniformly interpolated within each segment at each Δs = 0.1 units. 9 (a) shows the initial DDC used for the prostate image that is in 1 shown was estimated.

Table 1

To During initialization, the DDC is reshaped as detailed above described. The four points selected by the operator during the Initialization phase selected were, can be set before reshaping to prevent them from moving or you can without allowing them to be allowed to move with the rest of the DDC.

Even though the initialization routine of the present invention is capable is able to represent a wide range of prostate forms the initial form comes to lie near strong edges, the other Represent tissue or artifacts. This is what the DDC can and away from the desired one (Prostate) edge pushing, if the apparent forces by equations (4) and (5) can be used. The attraction of these edges can be reduced by using directional information, such as described in Worring M, Smeulders AWM, Staib IH, Duncan, JS, "Parameterized feasible Boundaries in gradient vector fields ", Computer Vision Image Understanding 1996; 63: 135-144. For US pictures the prostate this information represents the fact that the The inside of the prostate appears darker than the outside. Such information can be modified by modifying equation (5) to the vertex i of FIG DDC can be applied:

It is important to note that r ^ (x, y) is an outwardly pointing radial unit vector as defined by Equation 2. According to Equation 15, when the gray levels in the vicinity of the vertex i change from dark to light toward the outer radial vector, a force acts on the vertex; otherwise no force is applied. 9 (b) shows the DDC after forming.

During the Forming can affect the distance between adjacent vertices the DDC get bigger, which results in a worse representation of the prostate bone. After each iteration represented by equation (9), the DDC is resampled, leaving the distance between adjacent ones Keep vertices at a uniform value of 20 pixels remains, which was found to be a good representation of the Prostatorands ready. If no corner points are set, resampling can be done very easily. First, the x and y coordinates of each vertex are displayed in parametric form as in equation (10). However, it becomes the parametric variable s selected, as a length the curve from the first vertex to the point in question. The curve then becomes linear along its entire length at equidistant values Interpolation of the existing vertices of the length resampled. This results a new set of vertices used to represent the curve be while the old ones are discarded. This approach poses a problem when some vertices are set because the set ones Corner points can be eliminated after the interpolation. Around To avoid this, the set vertices become after the interpolation just back inserted into the DDC.

In extreme cases, the initialization routine can not place all of the vertices on the DDC near the actual prostatic border. 10 (a) shows an example where some portions of an initial DDC (indicated by the arrows) are removed from the prostatic border. As in 10 (b) As shown, this portion of the DDC does not converge on the desired edge because the apparent forces have a limited detection range. It is possible to correct the segmented edge by manipulating and reshaping the DDC. The operator may drag a corner point on the DDC to the desired edge. The vertex can then "set" so that it will not move when the DDC is reshaped. 10 (c) shows a case where three corner points (as represented by the squares) are drawn and set closer to the prostatic border. When a vertex is moved and set, the adjacent vertexes generally shift toward the fixed vertex under the influence of internal forces to minimize the local curvature. As these vertices shift into the detection area of the prostate bone, the apparent forces transport the vertices the rest of the way to the edge. The DDC is in after being reshaped 10 (d) shown.

The examples in 9 and 10 cases represent "small" or "medium" difficulty to be segmented. An example of greater challenge is in 11 shown where the prostate has a bulge above. This form is difficult to initialize with shape functions of cubic interpolation. One strategy to contour is to initialize the DDC by ignoring the bulge as in 11 (a) shown. The DDC does not reform into the bulge, as in 11 (b) shown. Upper vertices can, as in 11 (c) , pulled into position and fixed. The final segmented outline is in 11 (d) shown.

Other embodiments and variations are possible. For example, although the preferred embodiment set forth herein is addressed, the problem of prostate segmentation, the principle of the invention applied to edge segmentation of each fixed object be inclusive other organs within human or animal anatomy; suitable modification to the interpolation functions is ready used to initialize the edge before forming become.

It Can also be used for segmentation of objects, such as the prostate, in 3D be extended, the method according to the preferred embodiment is set forth as a 2D segmentation technique. Because the surface of the Organs is continuous, is a 2D image slice that only through small distances separated from the adjacent parallel slices is, by a similar Edge to the edges the adjacent discs characterized. Therefore, by a Object (that is Organ) cut into parallel slices with a sufficiently small distance will be similar to the edge of the object on adjacent 2D slices. Corresponding For example, the segmented edge in a slice may be an initial edge for the Segmentation process can be used in the adjacent disc. If the object edges enough are close, then leads the Using the starting edge in the adjacent discs too fast Segmentation without the requirement of editing.

The 3D embodiment of the present invention has been successfully implemented using a 3D ultrasound image of a patient's prostate. The 3D image was cut in parallel slices, 1 mm apart, to make 35 transaxial slices of the prostate. The approximately central disc was picked out as the starting disc (ie disc No. 15). The prostate in this disc was segmented by the technique described above using four starting points. This image did not require any processing and the resulting edge created by the segmentation is shown in Table 15. This edge was then used to initialize the segmentation of the adjacent disks 14 and 16. The resulting segmented edge (without machining) is shown in panels 14 and 16 of FIG 12 shown. The segmented edge in disk 14 was then used as the starting edge for disk 13 and the segmented edge in disk 16 was then used as the starting edge for disk 17. This process was continued until the segmentation of slice 1 and 34 was complete. The results of the whole process are in the tables of 12 shown.

Although the example presented above was based on parallel transaxial discs, one of ordinary skill in the art would prefer that the prostate (or any 3D object) be cut in any regular manner for this 3D segmentation procedure. For example, the prostate (or any 3D object) may be sliced into parallel slices in a coronal, sagital, or slanting direction, as well as in radial directions to make radial slices with an axis through the center of the prostate (or any other axis), such as in 13 shown. The radial cutting axis may be defined by any two opposite points used as starting points to initialize the segmentation in the central disk. The radial cutting can at equal angles, z. B. 1 degree, or varying angles depending on the complexity of the surface of the object.

It It is understood that the embodiments and modifications made within the scope and scope of the invention the defined attached claims lie.

Claims (11)

Method of object segmentation using an ultrasound image, comprising the steps: Operator selection a first plurality of starting points in the ultrasound image, the at least approximately lie on an edge of the object; Create an initial contour of the object by interpolation of additional core points, at least approximately lying on an edge of the object, using interpolation functions, which are selected for approximating the shape of the object; and automatic reshaping of at least the additional ones Key points, based on the edge data of the ultrasound image for Generation of a further multiplicity of points, which take the form of the Outline object. The method of claim 1, further comprising the steps of the interactive shifting of the further multiplicity of points and the automatic reshaping of the further plurality of points a final one Generate a variety of points that are a machined shape of the object outline. The method of claim 1 or 2, further comprising the steps of setting at least one of the first plurality of core points or of the further variety of points on theirs corresponding positions before the corresponding positions the steps of the automatic forming are achieved. The method of claim 1, 2 or 3, wherein each The step of automatically reshaping further comprises the step of converging of each point on marginal edges of the object using a Discrete Dynamic Contour Algorithm (DDC Algorithm) includes. The method of claim 1, 2 or 3, wherein the step of selecting Further, selecting the first one of the first plurality of key points and third points, the approximate lying on a first axis of symmetry of the object, and selecting second and fourth points, which approximate lie on a second axis of symmetry orthogonal to the first axis of symmetry is included. The method of claim 5, wherein the step of Generating an initial contour further comprising the step of (i) calculating corresponding coefficients of the interpolation functions from corresponding Positions of the first and third points, and the second and fourth Points, and (ii) the uniform Interpolieren of sections between the first and the second points, second and third points, third and third fourth points, and between the fourth and first points below Using the interpolation functions. The method of claim 6, wherein the step of uniform Interpolierenens further carrying out a Hermite interpolation of the sections in accordance with first and second cubic interpolation functions. The method of claim 4, wherein the step of Further, converging from each of the points to marginal edges of the object the step of modifying the Discrete Dynamic Contour Algorithm (DDC algorithm) includes to show apparent forces on the points only apply if the gray levels are adjacent to the corresponding ones the points from dark to light in a radial outward direction vary from each of the points. The method of claim 4, wherein the step of Further, converging from each of the points to marginal edges of the object the step of maintaining the respective intervals as the iteration progresses of the DDC algorithm to not less than 20 pixels between the corresponding points. The method of claim 1, further comprising Steps of selecting the further variety of points which outline the shape of the object in an image plane through the object, as corresponding to the additional Key points to which the step of automatic forming for an adjacent Parallel image plane through the object is applied to the shape to outline the object in three dimensions. An apparatus for object segmentation using an ultrasound image, comprising: means for obtaining an ultrasound image; Processing means for selecting a first plurality of key points in the ultrasound image that are at least approximately on an edge of the object, producing an initial contour of the object by interpolating additional kernel points that are at least approximately on an edge of the object; using interpolation functions selected to approximate the shape of the object; and automatically reforming at least the additional core points based on the edge data of the ultrasound image to produce a further plurality of points that outline the shape of the object.